xref: /openbmc/linux/net/ipv4/tcp_metrics.c (revision ebd09753)
1 // SPDX-License-Identifier: GPL-2.0
2 #include <linux/rcupdate.h>
3 #include <linux/spinlock.h>
4 #include <linux/jiffies.h>
5 #include <linux/module.h>
6 #include <linux/cache.h>
7 #include <linux/slab.h>
8 #include <linux/init.h>
9 #include <linux/tcp.h>
10 #include <linux/hash.h>
11 #include <linux/tcp_metrics.h>
12 #include <linux/vmalloc.h>
13 
14 #include <net/inet_connection_sock.h>
15 #include <net/net_namespace.h>
16 #include <net/request_sock.h>
17 #include <net/inetpeer.h>
18 #include <net/sock.h>
19 #include <net/ipv6.h>
20 #include <net/dst.h>
21 #include <net/tcp.h>
22 #include <net/genetlink.h>
23 
24 static struct tcp_metrics_block *__tcp_get_metrics(const struct inetpeer_addr *saddr,
25 						   const struct inetpeer_addr *daddr,
26 						   struct net *net, unsigned int hash);
27 
28 struct tcp_fastopen_metrics {
29 	u16	mss;
30 	u16	syn_loss:10,		/* Recurring Fast Open SYN losses */
31 		try_exp:2;		/* Request w/ exp. option (once) */
32 	unsigned long	last_syn_loss;	/* Last Fast Open SYN loss */
33 	struct	tcp_fastopen_cookie	cookie;
34 };
35 
36 /* TCP_METRIC_MAX includes 2 extra fields for userspace compatibility
37  * Kernel only stores RTT and RTTVAR in usec resolution
38  */
39 #define TCP_METRIC_MAX_KERNEL (TCP_METRIC_MAX - 2)
40 
41 struct tcp_metrics_block {
42 	struct tcp_metrics_block __rcu	*tcpm_next;
43 	possible_net_t			tcpm_net;
44 	struct inetpeer_addr		tcpm_saddr;
45 	struct inetpeer_addr		tcpm_daddr;
46 	unsigned long			tcpm_stamp;
47 	u32				tcpm_lock;
48 	u32				tcpm_vals[TCP_METRIC_MAX_KERNEL + 1];
49 	struct tcp_fastopen_metrics	tcpm_fastopen;
50 
51 	struct rcu_head			rcu_head;
52 };
53 
54 static inline struct net *tm_net(struct tcp_metrics_block *tm)
55 {
56 	return read_pnet(&tm->tcpm_net);
57 }
58 
59 static bool tcp_metric_locked(struct tcp_metrics_block *tm,
60 			      enum tcp_metric_index idx)
61 {
62 	return tm->tcpm_lock & (1 << idx);
63 }
64 
65 static u32 tcp_metric_get(struct tcp_metrics_block *tm,
66 			  enum tcp_metric_index idx)
67 {
68 	return tm->tcpm_vals[idx];
69 }
70 
71 static void tcp_metric_set(struct tcp_metrics_block *tm,
72 			   enum tcp_metric_index idx,
73 			   u32 val)
74 {
75 	tm->tcpm_vals[idx] = val;
76 }
77 
78 static bool addr_same(const struct inetpeer_addr *a,
79 		      const struct inetpeer_addr *b)
80 {
81 	return inetpeer_addr_cmp(a, b) == 0;
82 }
83 
84 struct tcpm_hash_bucket {
85 	struct tcp_metrics_block __rcu	*chain;
86 };
87 
88 static struct tcpm_hash_bucket	*tcp_metrics_hash __read_mostly;
89 static unsigned int		tcp_metrics_hash_log __read_mostly;
90 
91 static DEFINE_SPINLOCK(tcp_metrics_lock);
92 
93 static void tcpm_suck_dst(struct tcp_metrics_block *tm,
94 			  const struct dst_entry *dst,
95 			  bool fastopen_clear)
96 {
97 	u32 msval;
98 	u32 val;
99 
100 	tm->tcpm_stamp = jiffies;
101 
102 	val = 0;
103 	if (dst_metric_locked(dst, RTAX_RTT))
104 		val |= 1 << TCP_METRIC_RTT;
105 	if (dst_metric_locked(dst, RTAX_RTTVAR))
106 		val |= 1 << TCP_METRIC_RTTVAR;
107 	if (dst_metric_locked(dst, RTAX_SSTHRESH))
108 		val |= 1 << TCP_METRIC_SSTHRESH;
109 	if (dst_metric_locked(dst, RTAX_CWND))
110 		val |= 1 << TCP_METRIC_CWND;
111 	if (dst_metric_locked(dst, RTAX_REORDERING))
112 		val |= 1 << TCP_METRIC_REORDERING;
113 	tm->tcpm_lock = val;
114 
115 	msval = dst_metric_raw(dst, RTAX_RTT);
116 	tm->tcpm_vals[TCP_METRIC_RTT] = msval * USEC_PER_MSEC;
117 
118 	msval = dst_metric_raw(dst, RTAX_RTTVAR);
119 	tm->tcpm_vals[TCP_METRIC_RTTVAR] = msval * USEC_PER_MSEC;
120 	tm->tcpm_vals[TCP_METRIC_SSTHRESH] = dst_metric_raw(dst, RTAX_SSTHRESH);
121 	tm->tcpm_vals[TCP_METRIC_CWND] = dst_metric_raw(dst, RTAX_CWND);
122 	tm->tcpm_vals[TCP_METRIC_REORDERING] = dst_metric_raw(dst, RTAX_REORDERING);
123 	if (fastopen_clear) {
124 		tm->tcpm_fastopen.mss = 0;
125 		tm->tcpm_fastopen.syn_loss = 0;
126 		tm->tcpm_fastopen.try_exp = 0;
127 		tm->tcpm_fastopen.cookie.exp = false;
128 		tm->tcpm_fastopen.cookie.len = 0;
129 	}
130 }
131 
132 #define TCP_METRICS_TIMEOUT		(60 * 60 * HZ)
133 
134 static void tcpm_check_stamp(struct tcp_metrics_block *tm, struct dst_entry *dst)
135 {
136 	if (tm && unlikely(time_after(jiffies, tm->tcpm_stamp + TCP_METRICS_TIMEOUT)))
137 		tcpm_suck_dst(tm, dst, false);
138 }
139 
140 #define TCP_METRICS_RECLAIM_DEPTH	5
141 #define TCP_METRICS_RECLAIM_PTR		(struct tcp_metrics_block *) 0x1UL
142 
143 #define deref_locked(p)	\
144 	rcu_dereference_protected(p, lockdep_is_held(&tcp_metrics_lock))
145 
146 static struct tcp_metrics_block *tcpm_new(struct dst_entry *dst,
147 					  struct inetpeer_addr *saddr,
148 					  struct inetpeer_addr *daddr,
149 					  unsigned int hash)
150 {
151 	struct tcp_metrics_block *tm;
152 	struct net *net;
153 	bool reclaim = false;
154 
155 	spin_lock_bh(&tcp_metrics_lock);
156 	net = dev_net(dst->dev);
157 
158 	/* While waiting for the spin-lock the cache might have been populated
159 	 * with this entry and so we have to check again.
160 	 */
161 	tm = __tcp_get_metrics(saddr, daddr, net, hash);
162 	if (tm == TCP_METRICS_RECLAIM_PTR) {
163 		reclaim = true;
164 		tm = NULL;
165 	}
166 	if (tm) {
167 		tcpm_check_stamp(tm, dst);
168 		goto out_unlock;
169 	}
170 
171 	if (unlikely(reclaim)) {
172 		struct tcp_metrics_block *oldest;
173 
174 		oldest = deref_locked(tcp_metrics_hash[hash].chain);
175 		for (tm = deref_locked(oldest->tcpm_next); tm;
176 		     tm = deref_locked(tm->tcpm_next)) {
177 			if (time_before(tm->tcpm_stamp, oldest->tcpm_stamp))
178 				oldest = tm;
179 		}
180 		tm = oldest;
181 	} else {
182 		tm = kmalloc(sizeof(*tm), GFP_ATOMIC);
183 		if (!tm)
184 			goto out_unlock;
185 	}
186 	write_pnet(&tm->tcpm_net, net);
187 	tm->tcpm_saddr = *saddr;
188 	tm->tcpm_daddr = *daddr;
189 
190 	tcpm_suck_dst(tm, dst, true);
191 
192 	if (likely(!reclaim)) {
193 		tm->tcpm_next = tcp_metrics_hash[hash].chain;
194 		rcu_assign_pointer(tcp_metrics_hash[hash].chain, tm);
195 	}
196 
197 out_unlock:
198 	spin_unlock_bh(&tcp_metrics_lock);
199 	return tm;
200 }
201 
202 static struct tcp_metrics_block *tcp_get_encode(struct tcp_metrics_block *tm, int depth)
203 {
204 	if (tm)
205 		return tm;
206 	if (depth > TCP_METRICS_RECLAIM_DEPTH)
207 		return TCP_METRICS_RECLAIM_PTR;
208 	return NULL;
209 }
210 
211 static struct tcp_metrics_block *__tcp_get_metrics(const struct inetpeer_addr *saddr,
212 						   const struct inetpeer_addr *daddr,
213 						   struct net *net, unsigned int hash)
214 {
215 	struct tcp_metrics_block *tm;
216 	int depth = 0;
217 
218 	for (tm = rcu_dereference(tcp_metrics_hash[hash].chain); tm;
219 	     tm = rcu_dereference(tm->tcpm_next)) {
220 		if (addr_same(&tm->tcpm_saddr, saddr) &&
221 		    addr_same(&tm->tcpm_daddr, daddr) &&
222 		    net_eq(tm_net(tm), net))
223 			break;
224 		depth++;
225 	}
226 	return tcp_get_encode(tm, depth);
227 }
228 
229 static struct tcp_metrics_block *__tcp_get_metrics_req(struct request_sock *req,
230 						       struct dst_entry *dst)
231 {
232 	struct tcp_metrics_block *tm;
233 	struct inetpeer_addr saddr, daddr;
234 	unsigned int hash;
235 	struct net *net;
236 
237 	saddr.family = req->rsk_ops->family;
238 	daddr.family = req->rsk_ops->family;
239 	switch (daddr.family) {
240 	case AF_INET:
241 		inetpeer_set_addr_v4(&saddr, inet_rsk(req)->ir_loc_addr);
242 		inetpeer_set_addr_v4(&daddr, inet_rsk(req)->ir_rmt_addr);
243 		hash = ipv4_addr_hash(inet_rsk(req)->ir_rmt_addr);
244 		break;
245 #if IS_ENABLED(CONFIG_IPV6)
246 	case AF_INET6:
247 		inetpeer_set_addr_v6(&saddr, &inet_rsk(req)->ir_v6_loc_addr);
248 		inetpeer_set_addr_v6(&daddr, &inet_rsk(req)->ir_v6_rmt_addr);
249 		hash = ipv6_addr_hash(&inet_rsk(req)->ir_v6_rmt_addr);
250 		break;
251 #endif
252 	default:
253 		return NULL;
254 	}
255 
256 	net = dev_net(dst->dev);
257 	hash ^= net_hash_mix(net);
258 	hash = hash_32(hash, tcp_metrics_hash_log);
259 
260 	for (tm = rcu_dereference(tcp_metrics_hash[hash].chain); tm;
261 	     tm = rcu_dereference(tm->tcpm_next)) {
262 		if (addr_same(&tm->tcpm_saddr, &saddr) &&
263 		    addr_same(&tm->tcpm_daddr, &daddr) &&
264 		    net_eq(tm_net(tm), net))
265 			break;
266 	}
267 	tcpm_check_stamp(tm, dst);
268 	return tm;
269 }
270 
271 static struct tcp_metrics_block *tcp_get_metrics(struct sock *sk,
272 						 struct dst_entry *dst,
273 						 bool create)
274 {
275 	struct tcp_metrics_block *tm;
276 	struct inetpeer_addr saddr, daddr;
277 	unsigned int hash;
278 	struct net *net;
279 
280 	if (sk->sk_family == AF_INET) {
281 		inetpeer_set_addr_v4(&saddr, inet_sk(sk)->inet_saddr);
282 		inetpeer_set_addr_v4(&daddr, inet_sk(sk)->inet_daddr);
283 		hash = ipv4_addr_hash(inet_sk(sk)->inet_daddr);
284 	}
285 #if IS_ENABLED(CONFIG_IPV6)
286 	else if (sk->sk_family == AF_INET6) {
287 		if (ipv6_addr_v4mapped(&sk->sk_v6_daddr)) {
288 			inetpeer_set_addr_v4(&saddr, inet_sk(sk)->inet_saddr);
289 			inetpeer_set_addr_v4(&daddr, inet_sk(sk)->inet_daddr);
290 			hash = ipv4_addr_hash(inet_sk(sk)->inet_daddr);
291 		} else {
292 			inetpeer_set_addr_v6(&saddr, &sk->sk_v6_rcv_saddr);
293 			inetpeer_set_addr_v6(&daddr, &sk->sk_v6_daddr);
294 			hash = ipv6_addr_hash(&sk->sk_v6_daddr);
295 		}
296 	}
297 #endif
298 	else
299 		return NULL;
300 
301 	net = dev_net(dst->dev);
302 	hash ^= net_hash_mix(net);
303 	hash = hash_32(hash, tcp_metrics_hash_log);
304 
305 	tm = __tcp_get_metrics(&saddr, &daddr, net, hash);
306 	if (tm == TCP_METRICS_RECLAIM_PTR)
307 		tm = NULL;
308 	if (!tm && create)
309 		tm = tcpm_new(dst, &saddr, &daddr, hash);
310 	else
311 		tcpm_check_stamp(tm, dst);
312 
313 	return tm;
314 }
315 
316 /* Save metrics learned by this TCP session.  This function is called
317  * only, when TCP finishes successfully i.e. when it enters TIME-WAIT
318  * or goes from LAST-ACK to CLOSE.
319  */
320 void tcp_update_metrics(struct sock *sk)
321 {
322 	const struct inet_connection_sock *icsk = inet_csk(sk);
323 	struct dst_entry *dst = __sk_dst_get(sk);
324 	struct tcp_sock *tp = tcp_sk(sk);
325 	struct net *net = sock_net(sk);
326 	struct tcp_metrics_block *tm;
327 	unsigned long rtt;
328 	u32 val;
329 	int m;
330 
331 	sk_dst_confirm(sk);
332 	if (net->ipv4.sysctl_tcp_nometrics_save || !dst)
333 		return;
334 
335 	rcu_read_lock();
336 	if (icsk->icsk_backoff || !tp->srtt_us) {
337 		/* This session failed to estimate rtt. Why?
338 		 * Probably, no packets returned in time.  Reset our
339 		 * results.
340 		 */
341 		tm = tcp_get_metrics(sk, dst, false);
342 		if (tm && !tcp_metric_locked(tm, TCP_METRIC_RTT))
343 			tcp_metric_set(tm, TCP_METRIC_RTT, 0);
344 		goto out_unlock;
345 	} else
346 		tm = tcp_get_metrics(sk, dst, true);
347 
348 	if (!tm)
349 		goto out_unlock;
350 
351 	rtt = tcp_metric_get(tm, TCP_METRIC_RTT);
352 	m = rtt - tp->srtt_us;
353 
354 	/* If newly calculated rtt larger than stored one, store new
355 	 * one. Otherwise, use EWMA. Remember, rtt overestimation is
356 	 * always better than underestimation.
357 	 */
358 	if (!tcp_metric_locked(tm, TCP_METRIC_RTT)) {
359 		if (m <= 0)
360 			rtt = tp->srtt_us;
361 		else
362 			rtt -= (m >> 3);
363 		tcp_metric_set(tm, TCP_METRIC_RTT, rtt);
364 	}
365 
366 	if (!tcp_metric_locked(tm, TCP_METRIC_RTTVAR)) {
367 		unsigned long var;
368 
369 		if (m < 0)
370 			m = -m;
371 
372 		/* Scale deviation to rttvar fixed point */
373 		m >>= 1;
374 		if (m < tp->mdev_us)
375 			m = tp->mdev_us;
376 
377 		var = tcp_metric_get(tm, TCP_METRIC_RTTVAR);
378 		if (m >= var)
379 			var = m;
380 		else
381 			var -= (var - m) >> 2;
382 
383 		tcp_metric_set(tm, TCP_METRIC_RTTVAR, var);
384 	}
385 
386 	if (tcp_in_initial_slowstart(tp)) {
387 		/* Slow start still did not finish. */
388 		if (!tcp_metric_locked(tm, TCP_METRIC_SSTHRESH)) {
389 			val = tcp_metric_get(tm, TCP_METRIC_SSTHRESH);
390 			if (val && (tp->snd_cwnd >> 1) > val)
391 				tcp_metric_set(tm, TCP_METRIC_SSTHRESH,
392 					       tp->snd_cwnd >> 1);
393 		}
394 		if (!tcp_metric_locked(tm, TCP_METRIC_CWND)) {
395 			val = tcp_metric_get(tm, TCP_METRIC_CWND);
396 			if (tp->snd_cwnd > val)
397 				tcp_metric_set(tm, TCP_METRIC_CWND,
398 					       tp->snd_cwnd);
399 		}
400 	} else if (!tcp_in_slow_start(tp) &&
401 		   icsk->icsk_ca_state == TCP_CA_Open) {
402 		/* Cong. avoidance phase, cwnd is reliable. */
403 		if (!tcp_metric_locked(tm, TCP_METRIC_SSTHRESH))
404 			tcp_metric_set(tm, TCP_METRIC_SSTHRESH,
405 				       max(tp->snd_cwnd >> 1, tp->snd_ssthresh));
406 		if (!tcp_metric_locked(tm, TCP_METRIC_CWND)) {
407 			val = tcp_metric_get(tm, TCP_METRIC_CWND);
408 			tcp_metric_set(tm, TCP_METRIC_CWND, (val + tp->snd_cwnd) >> 1);
409 		}
410 	} else {
411 		/* Else slow start did not finish, cwnd is non-sense,
412 		 * ssthresh may be also invalid.
413 		 */
414 		if (!tcp_metric_locked(tm, TCP_METRIC_CWND)) {
415 			val = tcp_metric_get(tm, TCP_METRIC_CWND);
416 			tcp_metric_set(tm, TCP_METRIC_CWND,
417 				       (val + tp->snd_ssthresh) >> 1);
418 		}
419 		if (!tcp_metric_locked(tm, TCP_METRIC_SSTHRESH)) {
420 			val = tcp_metric_get(tm, TCP_METRIC_SSTHRESH);
421 			if (val && tp->snd_ssthresh > val)
422 				tcp_metric_set(tm, TCP_METRIC_SSTHRESH,
423 					       tp->snd_ssthresh);
424 		}
425 		if (!tcp_metric_locked(tm, TCP_METRIC_REORDERING)) {
426 			val = tcp_metric_get(tm, TCP_METRIC_REORDERING);
427 			if (val < tp->reordering &&
428 			    tp->reordering != net->ipv4.sysctl_tcp_reordering)
429 				tcp_metric_set(tm, TCP_METRIC_REORDERING,
430 					       tp->reordering);
431 		}
432 	}
433 	tm->tcpm_stamp = jiffies;
434 out_unlock:
435 	rcu_read_unlock();
436 }
437 
438 /* Initialize metrics on socket. */
439 
440 void tcp_init_metrics(struct sock *sk)
441 {
442 	struct dst_entry *dst = __sk_dst_get(sk);
443 	struct tcp_sock *tp = tcp_sk(sk);
444 	struct tcp_metrics_block *tm;
445 	u32 val, crtt = 0; /* cached RTT scaled by 8 */
446 
447 	sk_dst_confirm(sk);
448 	if (!dst)
449 		goto reset;
450 
451 	rcu_read_lock();
452 	tm = tcp_get_metrics(sk, dst, true);
453 	if (!tm) {
454 		rcu_read_unlock();
455 		goto reset;
456 	}
457 
458 	if (tcp_metric_locked(tm, TCP_METRIC_CWND))
459 		tp->snd_cwnd_clamp = tcp_metric_get(tm, TCP_METRIC_CWND);
460 
461 	val = tcp_metric_get(tm, TCP_METRIC_SSTHRESH);
462 	if (val) {
463 		tp->snd_ssthresh = val;
464 		if (tp->snd_ssthresh > tp->snd_cwnd_clamp)
465 			tp->snd_ssthresh = tp->snd_cwnd_clamp;
466 	} else {
467 		/* ssthresh may have been reduced unnecessarily during.
468 		 * 3WHS. Restore it back to its initial default.
469 		 */
470 		tp->snd_ssthresh = TCP_INFINITE_SSTHRESH;
471 	}
472 	val = tcp_metric_get(tm, TCP_METRIC_REORDERING);
473 	if (val && tp->reordering != val)
474 		tp->reordering = val;
475 
476 	crtt = tcp_metric_get(tm, TCP_METRIC_RTT);
477 	rcu_read_unlock();
478 reset:
479 	/* The initial RTT measurement from the SYN/SYN-ACK is not ideal
480 	 * to seed the RTO for later data packets because SYN packets are
481 	 * small. Use the per-dst cached values to seed the RTO but keep
482 	 * the RTT estimator variables intact (e.g., srtt, mdev, rttvar).
483 	 * Later the RTO will be updated immediately upon obtaining the first
484 	 * data RTT sample (tcp_rtt_estimator()). Hence the cached RTT only
485 	 * influences the first RTO but not later RTT estimation.
486 	 *
487 	 * But if RTT is not available from the SYN (due to retransmits or
488 	 * syn cookies) or the cache, force a conservative 3secs timeout.
489 	 *
490 	 * A bit of theory. RTT is time passed after "normal" sized packet
491 	 * is sent until it is ACKed. In normal circumstances sending small
492 	 * packets force peer to delay ACKs and calculation is correct too.
493 	 * The algorithm is adaptive and, provided we follow specs, it
494 	 * NEVER underestimate RTT. BUT! If peer tries to make some clever
495 	 * tricks sort of "quick acks" for time long enough to decrease RTT
496 	 * to low value, and then abruptly stops to do it and starts to delay
497 	 * ACKs, wait for troubles.
498 	 */
499 	if (crtt > tp->srtt_us) {
500 		/* Set RTO like tcp_rtt_estimator(), but from cached RTT. */
501 		crtt /= 8 * USEC_PER_SEC / HZ;
502 		inet_csk(sk)->icsk_rto = crtt + max(2 * crtt, tcp_rto_min(sk));
503 	} else if (tp->srtt_us == 0) {
504 		/* RFC6298: 5.7 We've failed to get a valid RTT sample from
505 		 * 3WHS. This is most likely due to retransmission,
506 		 * including spurious one. Reset the RTO back to 3secs
507 		 * from the more aggressive 1sec to avoid more spurious
508 		 * retransmission.
509 		 */
510 		tp->rttvar_us = jiffies_to_usecs(TCP_TIMEOUT_FALLBACK);
511 		tp->mdev_us = tp->mdev_max_us = tp->rttvar_us;
512 
513 		inet_csk(sk)->icsk_rto = TCP_TIMEOUT_FALLBACK;
514 	}
515 	/* Cut cwnd down to 1 per RFC5681 if SYN or SYN-ACK has been
516 	 * retransmitted. In light of RFC6298 more aggressive 1sec
517 	 * initRTO, we only reset cwnd when more than 1 SYN/SYN-ACK
518 	 * retransmission has occurred.
519 	 */
520 	if (tp->total_retrans > 1)
521 		tp->snd_cwnd = 1;
522 	else
523 		tp->snd_cwnd = tcp_init_cwnd(tp, dst);
524 	tp->snd_cwnd_stamp = tcp_jiffies32;
525 }
526 
527 bool tcp_peer_is_proven(struct request_sock *req, struct dst_entry *dst)
528 {
529 	struct tcp_metrics_block *tm;
530 	bool ret;
531 
532 	if (!dst)
533 		return false;
534 
535 	rcu_read_lock();
536 	tm = __tcp_get_metrics_req(req, dst);
537 	if (tm && tcp_metric_get(tm, TCP_METRIC_RTT))
538 		ret = true;
539 	else
540 		ret = false;
541 	rcu_read_unlock();
542 
543 	return ret;
544 }
545 
546 static DEFINE_SEQLOCK(fastopen_seqlock);
547 
548 void tcp_fastopen_cache_get(struct sock *sk, u16 *mss,
549 			    struct tcp_fastopen_cookie *cookie)
550 {
551 	struct tcp_metrics_block *tm;
552 
553 	rcu_read_lock();
554 	tm = tcp_get_metrics(sk, __sk_dst_get(sk), false);
555 	if (tm) {
556 		struct tcp_fastopen_metrics *tfom = &tm->tcpm_fastopen;
557 		unsigned int seq;
558 
559 		do {
560 			seq = read_seqbegin(&fastopen_seqlock);
561 			if (tfom->mss)
562 				*mss = tfom->mss;
563 			*cookie = tfom->cookie;
564 			if (cookie->len <= 0 && tfom->try_exp == 1)
565 				cookie->exp = true;
566 		} while (read_seqretry(&fastopen_seqlock, seq));
567 	}
568 	rcu_read_unlock();
569 }
570 
571 void tcp_fastopen_cache_set(struct sock *sk, u16 mss,
572 			    struct tcp_fastopen_cookie *cookie, bool syn_lost,
573 			    u16 try_exp)
574 {
575 	struct dst_entry *dst = __sk_dst_get(sk);
576 	struct tcp_metrics_block *tm;
577 
578 	if (!dst)
579 		return;
580 	rcu_read_lock();
581 	tm = tcp_get_metrics(sk, dst, true);
582 	if (tm) {
583 		struct tcp_fastopen_metrics *tfom = &tm->tcpm_fastopen;
584 
585 		write_seqlock_bh(&fastopen_seqlock);
586 		if (mss)
587 			tfom->mss = mss;
588 		if (cookie && cookie->len > 0)
589 			tfom->cookie = *cookie;
590 		else if (try_exp > tfom->try_exp &&
591 			 tfom->cookie.len <= 0 && !tfom->cookie.exp)
592 			tfom->try_exp = try_exp;
593 		if (syn_lost) {
594 			++tfom->syn_loss;
595 			tfom->last_syn_loss = jiffies;
596 		} else
597 			tfom->syn_loss = 0;
598 		write_sequnlock_bh(&fastopen_seqlock);
599 	}
600 	rcu_read_unlock();
601 }
602 
603 static struct genl_family tcp_metrics_nl_family;
604 
605 static const struct nla_policy tcp_metrics_nl_policy[TCP_METRICS_ATTR_MAX + 1] = {
606 	[TCP_METRICS_ATTR_ADDR_IPV4]	= { .type = NLA_U32, },
607 	[TCP_METRICS_ATTR_ADDR_IPV6]	= { .type = NLA_BINARY,
608 					    .len = sizeof(struct in6_addr), },
609 	/* Following attributes are not received for GET/DEL,
610 	 * we keep them for reference
611 	 */
612 #if 0
613 	[TCP_METRICS_ATTR_AGE]		= { .type = NLA_MSECS, },
614 	[TCP_METRICS_ATTR_TW_TSVAL]	= { .type = NLA_U32, },
615 	[TCP_METRICS_ATTR_TW_TS_STAMP]	= { .type = NLA_S32, },
616 	[TCP_METRICS_ATTR_VALS]		= { .type = NLA_NESTED, },
617 	[TCP_METRICS_ATTR_FOPEN_MSS]	= { .type = NLA_U16, },
618 	[TCP_METRICS_ATTR_FOPEN_SYN_DROPS]	= { .type = NLA_U16, },
619 	[TCP_METRICS_ATTR_FOPEN_SYN_DROP_TS]	= { .type = NLA_MSECS, },
620 	[TCP_METRICS_ATTR_FOPEN_COOKIE]	= { .type = NLA_BINARY,
621 					    .len = TCP_FASTOPEN_COOKIE_MAX, },
622 #endif
623 };
624 
625 /* Add attributes, caller cancels its header on failure */
626 static int tcp_metrics_fill_info(struct sk_buff *msg,
627 				 struct tcp_metrics_block *tm)
628 {
629 	struct nlattr *nest;
630 	int i;
631 
632 	switch (tm->tcpm_daddr.family) {
633 	case AF_INET:
634 		if (nla_put_in_addr(msg, TCP_METRICS_ATTR_ADDR_IPV4,
635 				    inetpeer_get_addr_v4(&tm->tcpm_daddr)) < 0)
636 			goto nla_put_failure;
637 		if (nla_put_in_addr(msg, TCP_METRICS_ATTR_SADDR_IPV4,
638 				    inetpeer_get_addr_v4(&tm->tcpm_saddr)) < 0)
639 			goto nla_put_failure;
640 		break;
641 	case AF_INET6:
642 		if (nla_put_in6_addr(msg, TCP_METRICS_ATTR_ADDR_IPV6,
643 				     inetpeer_get_addr_v6(&tm->tcpm_daddr)) < 0)
644 			goto nla_put_failure;
645 		if (nla_put_in6_addr(msg, TCP_METRICS_ATTR_SADDR_IPV6,
646 				     inetpeer_get_addr_v6(&tm->tcpm_saddr)) < 0)
647 			goto nla_put_failure;
648 		break;
649 	default:
650 		return -EAFNOSUPPORT;
651 	}
652 
653 	if (nla_put_msecs(msg, TCP_METRICS_ATTR_AGE,
654 			  jiffies - tm->tcpm_stamp,
655 			  TCP_METRICS_ATTR_PAD) < 0)
656 		goto nla_put_failure;
657 
658 	{
659 		int n = 0;
660 
661 		nest = nla_nest_start(msg, TCP_METRICS_ATTR_VALS);
662 		if (!nest)
663 			goto nla_put_failure;
664 		for (i = 0; i < TCP_METRIC_MAX_KERNEL + 1; i++) {
665 			u32 val = tm->tcpm_vals[i];
666 
667 			if (!val)
668 				continue;
669 			if (i == TCP_METRIC_RTT) {
670 				if (nla_put_u32(msg, TCP_METRIC_RTT_US + 1,
671 						val) < 0)
672 					goto nla_put_failure;
673 				n++;
674 				val = max(val / 1000, 1U);
675 			}
676 			if (i == TCP_METRIC_RTTVAR) {
677 				if (nla_put_u32(msg, TCP_METRIC_RTTVAR_US + 1,
678 						val) < 0)
679 					goto nla_put_failure;
680 				n++;
681 				val = max(val / 1000, 1U);
682 			}
683 			if (nla_put_u32(msg, i + 1, val) < 0)
684 				goto nla_put_failure;
685 			n++;
686 		}
687 		if (n)
688 			nla_nest_end(msg, nest);
689 		else
690 			nla_nest_cancel(msg, nest);
691 	}
692 
693 	{
694 		struct tcp_fastopen_metrics tfom_copy[1], *tfom;
695 		unsigned int seq;
696 
697 		do {
698 			seq = read_seqbegin(&fastopen_seqlock);
699 			tfom_copy[0] = tm->tcpm_fastopen;
700 		} while (read_seqretry(&fastopen_seqlock, seq));
701 
702 		tfom = tfom_copy;
703 		if (tfom->mss &&
704 		    nla_put_u16(msg, TCP_METRICS_ATTR_FOPEN_MSS,
705 				tfom->mss) < 0)
706 			goto nla_put_failure;
707 		if (tfom->syn_loss &&
708 		    (nla_put_u16(msg, TCP_METRICS_ATTR_FOPEN_SYN_DROPS,
709 				tfom->syn_loss) < 0 ||
710 		     nla_put_msecs(msg, TCP_METRICS_ATTR_FOPEN_SYN_DROP_TS,
711 				jiffies - tfom->last_syn_loss,
712 				TCP_METRICS_ATTR_PAD) < 0))
713 			goto nla_put_failure;
714 		if (tfom->cookie.len > 0 &&
715 		    nla_put(msg, TCP_METRICS_ATTR_FOPEN_COOKIE,
716 			    tfom->cookie.len, tfom->cookie.val) < 0)
717 			goto nla_put_failure;
718 	}
719 
720 	return 0;
721 
722 nla_put_failure:
723 	return -EMSGSIZE;
724 }
725 
726 static int tcp_metrics_dump_info(struct sk_buff *skb,
727 				 struct netlink_callback *cb,
728 				 struct tcp_metrics_block *tm)
729 {
730 	void *hdr;
731 
732 	hdr = genlmsg_put(skb, NETLINK_CB(cb->skb).portid, cb->nlh->nlmsg_seq,
733 			  &tcp_metrics_nl_family, NLM_F_MULTI,
734 			  TCP_METRICS_CMD_GET);
735 	if (!hdr)
736 		return -EMSGSIZE;
737 
738 	if (tcp_metrics_fill_info(skb, tm) < 0)
739 		goto nla_put_failure;
740 
741 	genlmsg_end(skb, hdr);
742 	return 0;
743 
744 nla_put_failure:
745 	genlmsg_cancel(skb, hdr);
746 	return -EMSGSIZE;
747 }
748 
749 static int tcp_metrics_nl_dump(struct sk_buff *skb,
750 			       struct netlink_callback *cb)
751 {
752 	struct net *net = sock_net(skb->sk);
753 	unsigned int max_rows = 1U << tcp_metrics_hash_log;
754 	unsigned int row, s_row = cb->args[0];
755 	int s_col = cb->args[1], col = s_col;
756 
757 	for (row = s_row; row < max_rows; row++, s_col = 0) {
758 		struct tcp_metrics_block *tm;
759 		struct tcpm_hash_bucket *hb = tcp_metrics_hash + row;
760 
761 		rcu_read_lock();
762 		for (col = 0, tm = rcu_dereference(hb->chain); tm;
763 		     tm = rcu_dereference(tm->tcpm_next), col++) {
764 			if (!net_eq(tm_net(tm), net))
765 				continue;
766 			if (col < s_col)
767 				continue;
768 			if (tcp_metrics_dump_info(skb, cb, tm) < 0) {
769 				rcu_read_unlock();
770 				goto done;
771 			}
772 		}
773 		rcu_read_unlock();
774 	}
775 
776 done:
777 	cb->args[0] = row;
778 	cb->args[1] = col;
779 	return skb->len;
780 }
781 
782 static int __parse_nl_addr(struct genl_info *info, struct inetpeer_addr *addr,
783 			   unsigned int *hash, int optional, int v4, int v6)
784 {
785 	struct nlattr *a;
786 
787 	a = info->attrs[v4];
788 	if (a) {
789 		inetpeer_set_addr_v4(addr, nla_get_in_addr(a));
790 		if (hash)
791 			*hash = ipv4_addr_hash(inetpeer_get_addr_v4(addr));
792 		return 0;
793 	}
794 	a = info->attrs[v6];
795 	if (a) {
796 		struct in6_addr in6;
797 
798 		if (nla_len(a) != sizeof(struct in6_addr))
799 			return -EINVAL;
800 		in6 = nla_get_in6_addr(a);
801 		inetpeer_set_addr_v6(addr, &in6);
802 		if (hash)
803 			*hash = ipv6_addr_hash(inetpeer_get_addr_v6(addr));
804 		return 0;
805 	}
806 	return optional ? 1 : -EAFNOSUPPORT;
807 }
808 
809 static int parse_nl_addr(struct genl_info *info, struct inetpeer_addr *addr,
810 			 unsigned int *hash, int optional)
811 {
812 	return __parse_nl_addr(info, addr, hash, optional,
813 			       TCP_METRICS_ATTR_ADDR_IPV4,
814 			       TCP_METRICS_ATTR_ADDR_IPV6);
815 }
816 
817 static int parse_nl_saddr(struct genl_info *info, struct inetpeer_addr *addr)
818 {
819 	return __parse_nl_addr(info, addr, NULL, 0,
820 			       TCP_METRICS_ATTR_SADDR_IPV4,
821 			       TCP_METRICS_ATTR_SADDR_IPV6);
822 }
823 
824 static int tcp_metrics_nl_cmd_get(struct sk_buff *skb, struct genl_info *info)
825 {
826 	struct tcp_metrics_block *tm;
827 	struct inetpeer_addr saddr, daddr;
828 	unsigned int hash;
829 	struct sk_buff *msg;
830 	struct net *net = genl_info_net(info);
831 	void *reply;
832 	int ret;
833 	bool src = true;
834 
835 	ret = parse_nl_addr(info, &daddr, &hash, 0);
836 	if (ret < 0)
837 		return ret;
838 
839 	ret = parse_nl_saddr(info, &saddr);
840 	if (ret < 0)
841 		src = false;
842 
843 	msg = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
844 	if (!msg)
845 		return -ENOMEM;
846 
847 	reply = genlmsg_put_reply(msg, info, &tcp_metrics_nl_family, 0,
848 				  info->genlhdr->cmd);
849 	if (!reply)
850 		goto nla_put_failure;
851 
852 	hash ^= net_hash_mix(net);
853 	hash = hash_32(hash, tcp_metrics_hash_log);
854 	ret = -ESRCH;
855 	rcu_read_lock();
856 	for (tm = rcu_dereference(tcp_metrics_hash[hash].chain); tm;
857 	     tm = rcu_dereference(tm->tcpm_next)) {
858 		if (addr_same(&tm->tcpm_daddr, &daddr) &&
859 		    (!src || addr_same(&tm->tcpm_saddr, &saddr)) &&
860 		    net_eq(tm_net(tm), net)) {
861 			ret = tcp_metrics_fill_info(msg, tm);
862 			break;
863 		}
864 	}
865 	rcu_read_unlock();
866 	if (ret < 0)
867 		goto out_free;
868 
869 	genlmsg_end(msg, reply);
870 	return genlmsg_reply(msg, info);
871 
872 nla_put_failure:
873 	ret = -EMSGSIZE;
874 
875 out_free:
876 	nlmsg_free(msg);
877 	return ret;
878 }
879 
880 static void tcp_metrics_flush_all(struct net *net)
881 {
882 	unsigned int max_rows = 1U << tcp_metrics_hash_log;
883 	struct tcpm_hash_bucket *hb = tcp_metrics_hash;
884 	struct tcp_metrics_block *tm;
885 	unsigned int row;
886 
887 	for (row = 0; row < max_rows; row++, hb++) {
888 		struct tcp_metrics_block __rcu **pp;
889 		bool match;
890 
891 		spin_lock_bh(&tcp_metrics_lock);
892 		pp = &hb->chain;
893 		for (tm = deref_locked(*pp); tm; tm = deref_locked(*pp)) {
894 			match = net ? net_eq(tm_net(tm), net) :
895 				!refcount_read(&tm_net(tm)->count);
896 			if (match) {
897 				*pp = tm->tcpm_next;
898 				kfree_rcu(tm, rcu_head);
899 			} else {
900 				pp = &tm->tcpm_next;
901 			}
902 		}
903 		spin_unlock_bh(&tcp_metrics_lock);
904 	}
905 }
906 
907 static int tcp_metrics_nl_cmd_del(struct sk_buff *skb, struct genl_info *info)
908 {
909 	struct tcpm_hash_bucket *hb;
910 	struct tcp_metrics_block *tm;
911 	struct tcp_metrics_block __rcu **pp;
912 	struct inetpeer_addr saddr, daddr;
913 	unsigned int hash;
914 	struct net *net = genl_info_net(info);
915 	int ret;
916 	bool src = true, found = false;
917 
918 	ret = parse_nl_addr(info, &daddr, &hash, 1);
919 	if (ret < 0)
920 		return ret;
921 	if (ret > 0) {
922 		tcp_metrics_flush_all(net);
923 		return 0;
924 	}
925 	ret = parse_nl_saddr(info, &saddr);
926 	if (ret < 0)
927 		src = false;
928 
929 	hash ^= net_hash_mix(net);
930 	hash = hash_32(hash, tcp_metrics_hash_log);
931 	hb = tcp_metrics_hash + hash;
932 	pp = &hb->chain;
933 	spin_lock_bh(&tcp_metrics_lock);
934 	for (tm = deref_locked(*pp); tm; tm = deref_locked(*pp)) {
935 		if (addr_same(&tm->tcpm_daddr, &daddr) &&
936 		    (!src || addr_same(&tm->tcpm_saddr, &saddr)) &&
937 		    net_eq(tm_net(tm), net)) {
938 			*pp = tm->tcpm_next;
939 			kfree_rcu(tm, rcu_head);
940 			found = true;
941 		} else {
942 			pp = &tm->tcpm_next;
943 		}
944 	}
945 	spin_unlock_bh(&tcp_metrics_lock);
946 	if (!found)
947 		return -ESRCH;
948 	return 0;
949 }
950 
951 static const struct genl_ops tcp_metrics_nl_ops[] = {
952 	{
953 		.cmd = TCP_METRICS_CMD_GET,
954 		.doit = tcp_metrics_nl_cmd_get,
955 		.dumpit = tcp_metrics_nl_dump,
956 		.policy = tcp_metrics_nl_policy,
957 	},
958 	{
959 		.cmd = TCP_METRICS_CMD_DEL,
960 		.doit = tcp_metrics_nl_cmd_del,
961 		.policy = tcp_metrics_nl_policy,
962 		.flags = GENL_ADMIN_PERM,
963 	},
964 };
965 
966 static struct genl_family tcp_metrics_nl_family __ro_after_init = {
967 	.hdrsize	= 0,
968 	.name		= TCP_METRICS_GENL_NAME,
969 	.version	= TCP_METRICS_GENL_VERSION,
970 	.maxattr	= TCP_METRICS_ATTR_MAX,
971 	.netnsok	= true,
972 	.module		= THIS_MODULE,
973 	.ops		= tcp_metrics_nl_ops,
974 	.n_ops		= ARRAY_SIZE(tcp_metrics_nl_ops),
975 };
976 
977 static unsigned int tcpmhash_entries;
978 static int __init set_tcpmhash_entries(char *str)
979 {
980 	ssize_t ret;
981 
982 	if (!str)
983 		return 0;
984 
985 	ret = kstrtouint(str, 0, &tcpmhash_entries);
986 	if (ret)
987 		return 0;
988 
989 	return 1;
990 }
991 __setup("tcpmhash_entries=", set_tcpmhash_entries);
992 
993 static int __net_init tcp_net_metrics_init(struct net *net)
994 {
995 	size_t size;
996 	unsigned int slots;
997 
998 	if (!net_eq(net, &init_net))
999 		return 0;
1000 
1001 	slots = tcpmhash_entries;
1002 	if (!slots) {
1003 		if (totalram_pages >= 128 * 1024)
1004 			slots = 16 * 1024;
1005 		else
1006 			slots = 8 * 1024;
1007 	}
1008 
1009 	tcp_metrics_hash_log = order_base_2(slots);
1010 	size = sizeof(struct tcpm_hash_bucket) << tcp_metrics_hash_log;
1011 
1012 	tcp_metrics_hash = kvzalloc(size, GFP_KERNEL);
1013 	if (!tcp_metrics_hash)
1014 		return -ENOMEM;
1015 
1016 	return 0;
1017 }
1018 
1019 static void __net_exit tcp_net_metrics_exit_batch(struct list_head *net_exit_list)
1020 {
1021 	tcp_metrics_flush_all(NULL);
1022 }
1023 
1024 static __net_initdata struct pernet_operations tcp_net_metrics_ops = {
1025 	.init		=	tcp_net_metrics_init,
1026 	.exit_batch	=	tcp_net_metrics_exit_batch,
1027 };
1028 
1029 void __init tcp_metrics_init(void)
1030 {
1031 	int ret;
1032 
1033 	ret = register_pernet_subsys(&tcp_net_metrics_ops);
1034 	if (ret < 0)
1035 		panic("Could not allocate the tcp_metrics hash table\n");
1036 
1037 	ret = genl_register_family(&tcp_metrics_nl_family);
1038 	if (ret < 0)
1039 		panic("Could not register tcp_metrics generic netlink\n");
1040 }
1041